Improve Xenium performance, fix multinucleate cells bug

[LucaMarconato]

The last commit message is indicative of this PR. Copy-pasting it below.

Refactor xenium polygon reader to use zarr-based indices mapping
Replace slow parquet string-based cell_id grouping with zarr polygon_sets
for both nucleus and cell boundaries. This fixes a multinucleate cell bug
where multiple nuclei sharing the same cell_id were incorrectly merged
into a single polygon, and improves performance by avoiding expensive
string operations on large parquet columns.

Key changes:

• Split _get_labels_and_indices_mapping into focused functions:
  _get_labels, _get_indices_mapping_from_zarr, _get_indices_mapping_legacy
• Nucleus boundaries now use integer label_index as GeoDataFrame index
  (with cell_id as a column), correctly handling multinucleate cells
• Cell boundaries keep string cell_id as GeoDataFrame index (legacy)
• Read only needed parquet columns for faster I/O
• Use integer label_id for fast change detection when available